Post-merge follow-up: EEGLAB test-port stack (#329-#372)

[arnodelorme]

🤖 Filed by an agent at @adelorme's request, at merge time for the EEGLAB test-port stack (#329–#372).

Umbrella issue recording what was and was not examined before the stack was merged, and tracking the follow-up work.

What was merged

43 linear PRs (#329–#370, #372; #371 closed as an abandoned sibling), 57 commits, 387 files, +37,673 / −866, merged with zero divergence from develop @ 1f4c74f0.

Surface	Files	Lines
New modules	141	+14,911
Modified modules	114	+4,843 / −625
New/changed tests	99	+16,519
Docs	29	+1,017 / −99
docs/source/changelog.rst	1	+203

Modified source by subsystem: studyfunc 16 files (+2,358/−214), popfunc 35 (+1,298/−253), adminfunc 3 (+368/−19), sigprocfunc 15 (+214/−33), resources 27 (+183/−48), top-level __init__.py (+140), guifunc 6 (+105/−34), statistics 4 (+69/−11), miscfunc 3 (+58/−6), plugins/firfilt 1 (+44/−4), timefreqfunc 2 (+3/−3), plugins/clean_rawdata 1 (+3).

What was verified before merging

• Tip CI green on the full matrix (Python 3.10/3.11/3.12 ubuntu, 3.12 macOS + Windows, ruff+ty, pre-commit, docs), run 2026-09-14. Merge-base was develop HEAD at merge time, so that run tested exactly the tree that landed.
• External parity harness (parity_repo): EEGPrep vs live MATLAB/EEGLAB through eeglabcompat, 13 subjects × 7 pipeline stages. Import and re-referencing exactly identical; ASR/ICA ≤ 0.000014 µV; two MATLAB R2026a ICLabel outliers reproduced identically on unmodified develop, confirming they predate this work. The harness drives 18 entry points, 8 of which this stack modifies: pop_loadset, pop_saveset, pop_select, pop_reref, pop_epoch, pop_rmbase, eeg_interp, eeg_picard. See Record parity harness provenance and confirm ASR rejection mode #392 for provenance and one open question about ASR rejection mode.
• Individual code checks recorded in Restore two pop function tests removed by the test-port stack #388–Audit public API surface and capability claims from the test-port stack #393.

What was NOT reviewed

Stated plainly so future debugging starts from an accurate picture. No line-by-line review of the source was performed before merging, and the parity harness cannot reach the following, which together account for the large majority of the changed files:

Subsystem	Changed files	Why not covered
studyfunc (STUDY, LIMO, clustering, measure plots)	30	No STUDY stage in the pipeline
Binary readers/writers (CNT, BrainVision, ERPSS, EGI RAW/MAT, GDF, BDF, neurodat)	22	Harness loads BIDS/.set only
miscfunc	53	Not in the pipeline
guifunc	10	Not in the pipeline
statistics	5	Not in the pipeline

Also outside harness reach: apply_eegfilt_legacy / _legacy_firls (reached only from pop_eegfilt and sigprocfunc/eegfilt; clean_artifacts uses its own filtering) and _cluster_kmeans / kmeans_st.

Additional context: of the 99 new/changed test files, 0 reference matlab, octave, eeglabcompat, or pytest.mark.parity — they are Python self-consistency tests. PRs #329–#331 ran 5 CI checks rather than 9, so the bottom of the stack never ran the test suite on its own.

Follow-up

• Restore two pop function tests removed by the test-port stack #388 — restore two removed pop function tests
• Verify STUDY k-means canonicalization does not shift cluster numbering #389 — verify STUDY k-means canonicalization does not shift cluster numbering
• Quantify the legacy eegfilt filtfilt padlen change #390 — quantify the legacy eegfilt filtfilt padlen change
• Validate the six new binary reader families against real recordings #391 — validate the six new binary reader families against real recordings
• Record parity harness provenance and confirm ASR rejection mode #392 — record parity harness provenance and confirm ASR rejection mode
• Audit public API surface and capability claims from the test-port stack #393 — audit public API surface and capability claims

#391 carries the highest residual risk: new code, outside the harness path, tested only against fixtures generated by the same commits that read them.

Note on PR titles

26 of the 42 content PRs add more source lines than test lines under "Port … tests" titles — #347 adds 1,404 source lines across 35 files under "Port current EEGLAB numerical utility tests". Titles that name the dominant content would let reviewers allocate attention correctly at triage time. Several PRs did this well (#363 "Implement CORRMAP component matching", #364 "Add standalone FieldTrip-style statistics", #369 "Implement standalone LIMO workflows").